Xberg Document Extraction
AdoptFactorial rates Xberg as Adopt (2026). In Subtool, Xberg runs as a stateless sidecar (xberg serve) that extracts every uploaded tender document before it is chunked and stored in pgvector.
Xberg is an open-source document extraction engine and the successor to Kreuzberg. It turns PDFs, Office documents, images and many other formats into text, tables and metadata, with OCR (for example Tesseract) for scanned pages. It can be embedded as a library in several languages, used from the command line, or run as a REST API server (xberg serve) or MCP server.
It covers the first step of a retrieval pipeline (extract, chunk, embed, store) without sending documents to a hosted service.
Updates
In Subtool, Xberg runs as a stateless sidecar (xberg serve) that extracts every uploaded tender document before it is chunked and stored in pgvector. We call it over HTTP rather than through language bindings, to keep our Go backend free of cgo, and pin the image version. Two lessons: OCR needs the German language model configured, or umlauts come out mangled, and no OCR recovers table structure from scanned price sheets, so we mark such extractions with a warning instead of treating the figures as reliable.